Fix some warnings in mlas (#10034)

onnxruntime/core/mlas/lib/convolve.cpp

@@ -1016,7 +1016,10 @@ Return Value:
         }
     }
 }
-
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(push)
+#pragma warning(disable : 26451)
+#endif
 void
 MLASCALL
 MlasConvPrepare(
@@ -1289,3 +1292,6 @@ Return Value:
         *WorkingBufferSize = TargetThreadCount * MLAS_CONV_WORKING_BUFFER_SIZE_PER_THREAD;
     }
 }
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(pop)
+#endif

onnxruntime/core/mlas/lib/dgemm.cpp

@@ -803,7 +803,10 @@ Return Value:
         Data->alpha, A, lda, B, ldb, Data->beta, C, ldc);
 }
 
-
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(push)
+#pragma warning(disable : 26451)
+#endif
 void
 MLASCALL
 MlasGemmBatch(
@@ -880,5 +883,7 @@ MlasGemmBatch(
     });
 
 }
-
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(pop)
+#endif
 #endif

onnxruntime/core/mlas/lib/intrinsics/avx2/qladd_avx2.cpp

@@ -185,7 +185,7 @@ MlasQLinearAddKernelAvx2Helper(
         if (k > 0) {
             const __m256i mask = _mm256_cmpgt_epi32(_mm256_set1_epi32(k), _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0));
             _mm256_maskstore_epi32((int*)OutputC, mask, vc);
-            OutputC += k * 4;
+            OutputC += static_cast<ptrdiff_t>(k) * 4;
         }
 
         int r = static_cast<int>(n % 4);

onnxruntime/core/mlas/lib/mlasi.h

@@ -933,6 +933,11 @@ MlasPartitionWork(
 //
 // Helpers to cast a floating point type to and from an integer bit format.
 //
+#if defined(_MSC_VER) && !defined(__clang__)
+  #pragma warning(push)
+  // VC++ suggests we can attempt to make 'MlasBitsOfFp32' constexpr, but it is not valid.
+  #pragma warning(disable:26497) 
+#endif
 
 MLAS_FORCEINLINE
 uint32_t
@@ -961,7 +966,9 @@ MlasFp32FromBits(
     u.IntegerValue = IntegerValue;
     return u.FloatValue;
 }
-
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(pop)
+#endif
 
 #if defined(MLAS_TARGET_WASM_SCALAR)
 

onnxruntime/core/mlas/lib/pooling.cpp

@@ -66,7 +66,7 @@ void
 
 struct MLAS_MAXIMUM_POOLING
 {
-    static float InitialValue()
+    static constexpr float InitialValue()
     {
         return std::numeric_limits<float>::lowest();
     }
@@ -76,7 +76,7 @@ struct MLAS_MAXIMUM_POOLING
         return MlasBroadcastFloat32x4(InitialValue());
     }
 
-    static float Reduce(float Reduction, float Value)
+    static constexpr float Reduce(float Reduction, float Value)
     {
         return std::max(Reduction, Value);
     }
@@ -91,7 +91,7 @@ struct MLAS_MAXIMUM_POOLING
         return MlasReduceMaximumFloat32x4(Reduction);
     }
 
-    static float AveragePool(float Reduction, float Size)
+    static constexpr float AveragePool(float Reduction, float Size)
     {
         MLAS_UNREFERENCED_PARAMETER(Size);
 
@@ -147,7 +147,7 @@ struct MLAS_AVERAGE_POOLING
         return MlasZeroFloat32x4();
     }
 
-    static float Reduce(float Reduction, float Value)
+    static constexpr float Reduce(float Reduction, float Value)
     {
         return Reduction + Value;
     }
@@ -162,7 +162,7 @@ struct MLAS_AVERAGE_POOLING
         return MlasReduceAddFloat32x4(Reduction);
     }
 
-    static float AveragePool(float Reduction, float Size)
+    static constexpr float AveragePool(float Reduction, float Size)
     {
         return Reduction / Size;
     }

onnxruntime/core/mlas/lib/qgemm.cpp

@@ -139,7 +139,11 @@ Return Value:
 {
     MlasGemmBatch(Shape, &DataParams, 1, ThreadPool);
 }
-
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(push)
+// VC++ suggests we can attempt to make 'MlasBitsOfFp32' constexpr, but it is not valid.
+#pragma warning(disable : 26451)
+#endif
 void
 MLASCALL
 MlasGemmBatch(
@@ -216,7 +220,9 @@ MlasGemmBatch(
         MlasGemmQuantThreaded(&WorkBlock, &Shape, &DataParams[gemm_i], blk_i);
     });
 }
-
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(pop)
+#endif
 
 size_t
 MLASCALL

onnxruntime/core/mlas/lib/qgemm.h

@@ -73,7 +73,7 @@ MlasGemmQuantTryGemvKernel(
 }
 
 template <typename KernelType>
-MLAS_FORCEINLINE
+MLAS_FORCEINLINE constexpr
 int32_t
 MlasGemmQuantFixupZeroPointA(
     int32_t ZeroPointA,
@@ -84,7 +84,7 @@ MlasGemmQuantFixupZeroPointA(
 }
 
 template<typename KernelType>
-int32_t
+int32_t constexpr
 MlasGemmQuantFixupZeroPointB(
     int32_t ZeroPointB,
     bool BIsSigned

onnxruntime/core/mlas/lib/qgemm_kernel_avx2.cpp

@@ -115,7 +115,7 @@ MlasGemmQuantTryGemvKernel<MLAS_GEMM_U8S8_KERNEL_AVX2>(
 }
 
 template<>
-MLAS_FORCEINLINE
+MLAS_FORCEINLINE constexpr
 int32_t
 MlasGemmQuantFixupZeroPointB<MLAS_GEMM_U8S8_KERNEL_AVX2>(
     int32_t ZeroPointB,

onnxruntime/core/mlas/lib/qgemm_kernel_default.cpp

@@ -34,7 +34,7 @@ constexpr MLAS_GEMM_QUANT_STRIDES MLAS_GEMM_QUANT_KERNEL_DEFAULT::Strides;
 constexpr MLAS_GEMM_QUANT_STRIDES MLAS_GEMM_QUANT_KERNEL_DEFAULT::PackedStrides;
 
 template<>
-MLAS_FORCEINLINE
+MLAS_FORCEINLINE constexpr 
 int32_t
 MlasGemmQuantFixupZeroPointA<MLAS_GEMM_QUANT_KERNEL_DEFAULT>(
     int32_t ZeroPointA,

onnxruntime/core/mlas/lib/qgemm_kernel_sse.cpp

@@ -32,7 +32,7 @@ constexpr size_t MLAS_GEMM_U8X8_KERNEL_SSE::PackedK;
 constexpr MLAS_GEMM_QUANT_STRIDES MLAS_GEMM_U8X8_KERNEL_SSE::Strides;
 
 template<>
-MLAS_FORCEINLINE
+MLAS_FORCEINLINE constexpr
 int32_t
 MlasGemmQuantFixupZeroPointB<MLAS_GEMM_U8X8_KERNEL_SSE>(
     int32_t ZeroPointB,

onnxruntime/core/mlas/lib/qlgavgpool.cpp

@@ -624,28 +624,37 @@ MlasQLinearGlobalAveragePoolNhwcSingleBatch(
 #if defined(MLAS_TARGET_IX86)
         switch (ImageSize) {
             case 1:
-                i1 = ZeroBuffer; /* fall through */
+                i1 = ZeroBuffer;
+                [[fallthrough]];
             case 2:
-                i2 = ZeroBuffer; /* fall through */
+                i2 = ZeroBuffer;
+                [[fallthrough]];
             case 3:
-                i3 = ZeroBuffer; /* fall through */
+                i3 = ZeroBuffer;
+                [[fallthrough]];
             default:
                 break;
         }
 #else
         switch (ImageSize) {
             case 1:
-                i1 = ZeroBuffer; /* fall through */
+                i1 = ZeroBuffer;
+                [[fallthrough]];
             case 2:
-                i2 = ZeroBuffer; /* fall through */
+                i2 = ZeroBuffer;
+                [[fallthrough]];
             case 3:
-                i3 = ZeroBuffer; /* fall through */
+                i3 = ZeroBuffer;
+                [[fallthrough]];
             case 4:
-                i4 = ZeroBuffer; /* fall through */
+                i4 = ZeroBuffer;
+                [[fallthrough]];
             case 5:
-                i5 = ZeroBuffer; /* fall through */
+                i5 = ZeroBuffer;
+                [[fallthrough]];
             case 6:
-                i6 = ZeroBuffer; /* fall through */
+                i6 = ZeroBuffer;
+                [[fallthrough]];
             default:
                 break;
         }

onnxruntime/core/mlas/lib/sgemm.cpp

@@ -1552,7 +1552,10 @@ Return Value:
             DataParams->alpha, A, lda, B, ldb, DataParams->beta, C, ldc);
     }
 }
-
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(push)
+#pragma warning(disable : 26451)
+#endif
 void
 MLASCALL
 MlasGemmBatch(
@@ -1631,6 +1634,9 @@ MlasGemmBatch(
             TransA, TransB, M, N, K, &(Data[GemmIdx]), ThreadIdx);
     });
 }
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma warning(pop)
+#endif
 
 size_t
 MLASCALL